Signals

Signals are what the team currently believes — each one citable as SIG-{number}.

How it works

A signal answers three independent questions, and no single field mixes them:

• stage — how much evidence stands behind it: emerging (still forming) or established (multiple studies, multiple participants). Computed from evidence counts, never from anyone's judgment.
• actionable — is a recommendation on the table? True while the signal has a published recommendation that is neither dismissed nor validated (accepted, shipped, or stale all keep it actionable); never true once archived. Derived, not stored.
• archived — is the loop closed? archiveReason says how: validated (you acted and confirmed the outcome) or abandoned (set aside).

An emerging signal can be actionable — evidence maturity and "worth acting on" are different questions. Alongside the axes, a signal carries a trajectory (rising, stable, declining, new) and a confidence band (high, moderate, low). Analysis reserves high for beliefs corroborated across multiple consistent studies; a pattern from one or two participants stays low no matter how good the quotes are.

After a linked recommendation ships, the signal may also carry a postShipReceipt — a dated record that the quiet window passed with real listening behind it: shipped when, no new unprompted corroboration for how many days, across how many interviews, trajectory as observed. Quiet with nobody interviewed proves nothing, so the receipt only stamps once at least two people were interviewed inside the quiet span. It is an observation, not a verdict; whether the fix truly landed stays your call.

Themes are not signals

Don't mix their IDs: theme IDs feed study and evidence tools, signal IDs feed signal tools.

What you can do

Everything analysis produces, you can overrule. In the app, archive a signal you consider settled or wrong — and restore it later — or lock its wording so analysis stops rewriting it. Corrections are first-class inputs, not fights with the system; the API and MCP tools read signals but don't carry these correction verbs.

When it's quiet

Neither an empty inbox nor a quiet signal board alone proves there's nothing to learn. It may only mean you haven't interviewed the people who would tell you. Research gaps keep that distinction visible.

From your agent

• list_signals — every belief with stage, trajectory, and confidence; filter by state.
• get_signal — one signal's full packet: memo, evidence preview, per-study breakdown.
• get_evidence — the verbatim quotes behind it (pass signalId).